text summarization
/ TEKST sum-uh-rih-ZAY-shun /
Text summarization is software that takes a long document and boils it down to a short version that keeps the gist — the same thing a good study aid does when it condenses a chapter into a few sentences. The goal is to save you reading time without losing what matters: the main points survive, the padding falls away.
There are two fundamentally different ways to do it, and the difference matters. Extractive summarization picks out the most important sentences from the original and stitches them together verbatim — safe, since every word is genuine, but sometimes choppy. Abstractive summarization instead writes fresh sentences in its own words, the way a person would paraphrase — smoother and more natural, but riskier, because a model writing new text can also write things the source never said. Most modern systems are abstractive, built on the same next-word-prediction machinery as other language models.
Summarization shows up in news digests, meeting-note tools, search snippets, and "too long; didn't read" buttons. The candid warning is about abstractive systems: they can hallucinate — confidently inserting a fact, name, or number that isn't in the original, or quietly flipping a conclusion. A summary that reads beautifully can still be unfaithful to its source, which is exactly why a quick human check still matters when the stakes are real.
From a 600-word article on a flood, extractive output might lift the two sentences with the death toll and the cause. Abstractive output writes one fresh sentence: "Heavy rain caused a river to burst its banks, killing twelve." — smoother, but it must be checked against the source.
Extractive lifts real sentences; abstractive writes new ones — and can drift from the truth.
Abstractive summaries can hallucinate facts that were never in the source, and a fluent summary is not automatically a faithful one. ROUGE measures overlap with a reference, not factual correctness.